Does "train" already include train/val/test division if a feedforward neural network or should I do it manually before?

Hello everyone! I have the following script:

-----------------------

inputs=training_set'; %training_set is my matrix of features, with observation on rows and features on columns

net = feedforwardnet(10); %I set 10 hidden layers and by default trains with "trainlm" function

[net,tr] = train(net,inputs,new_training_labels');

y_predict=net(inputs);

----------------------

My question is: does train already divide "inputs" in train/val/test and runs the crossvalidation automatically or should I do it before and then write something like:

y_predict=net(inputs(trainInd,:));

Thank you for your time!

Answer 1

[Greg Heath]

1. Use fitnet for regression and patternnet for classification.

2. Use the help and doc commands for sample code.

3. Run the sample code. If you have problems, here we are!

4. Adjust your code until you are satisfied.

5. Try your data with the code

etc

 [ net tr y e ] = train( net, x, t );
 % y = net(x); e = t-y;
 NMSE = mse(e)/mse(t-mean(t,2)) % NORMALIZED MSE

All of the datadivision details are in the training record tr

To watch all of the goodies tumble out just type

 tr = tr

Hope this helps.

Thank you for formally accepting my answer

Greg

Answer 2

[Helper]

Yes, it already includes train, validation, and test.

The default training function of “feedforwardnet” is “trainlm”. If you check the documentation of “ trainlm ”.

You will find “trainlm supports training with validation and test vectors if the network's NET.divideFcn property is set to a data division function.” Then if you check the “divideFcn” of “net” in your case, it would be “dividerand” (in MATLAB R2017a).

For more information of “dividerand”, please type “help nndivision”, then you would get a more detailed explanation.

[Greg Heath]

For more information on any entity FIRST use the help and doc commands (in that order!). The recent versions of the doc commands are not as helpful as the older versions.

 help dividerand

 doc dividerand

Hope this helps

Greg